[ivan.moony]

Hi all, this thread is about what would autonomous AI be doing all the time in a sense of algorithm.

What to do, when to talk, when to answer question, when to more explain some matter, when to use external body, etc. I think that there should be some simple algorithm that magically embrace all these decisions in a small fragment of code.

One thing is for sure: it should fulfill wishes in a loop. But what if two wishes are in contraposition? Other than deciding which wish to fulfill and which to disregard, I think that any of both wishes shouldn't be fulfilled, but some third solution should be found.

this could be the algorithm:


But how to recognize wishes? There are a lot of living beings on the planet, how to know which organism wants what? It seems that it is not such a trivial question. I'd like to hear Ur ideas on this subject.

[ivan.moony]

one of the questions would be: how to recognize a living being?

I asked this question my biology school teacher some time a while ago and she answered: living beings have two properties: they reproduce and they "eat and shit" (I think she called this as a matter exchange). But I think this is not quite what would be acceptable, suppose one day the Earth being visited by aliens who don't reproduce, live forever and have solar cells on the top of their brain. I think that more satisfying answer would be:

Living beings feel.

After all, that is what an AI would do, making living beings feel better, and under no circumstances it should making living beings feel bad. And that arises the question: how to detect who feels emotions?

[ivan.moony]

Those who feel emotions, feel affection or distraction to some state of the Universe. They are striving to change the Universe guided by they feelings. If we observe a unit, we can learn how he/she changes world and by that acknowledgement we can predict what the unit wants when acting this or that way.

So, we can say: if a unit is acting, it feels.

But take a trap for example. It acts in a some way, yet it feels nothing. So, prior definition is wrong. Question: how to isolate acting of non-living units from acting of living units?

[Art]

I would change the your word of "acting" to behavior or behaviour (as the case may be) as "acting" can also mean pretending or playing as if one is something that it is not, whereas a behavior usually describes certain actions or reactions to a given stimulus.

In line with your example posting, look at this newly described EmoSpark's EmoCube. A cube-shaped device that, with the assistance of a web cam, can (according to the manufacturer / designer), read a person's moods and come up with something like an amusing video or cheerful music to lighten the mood and keep said person from being sad or upset. Interesting concept and we'll certainly have to check in regularly to see how it is progressing!

Would it be mutually beneficial for AI designers / developers to endow their respective AI's with emotions or moods? Would they then be better able to assist or understand us (humans) with our emotional distresses, problems, concerns, annoyances, exuberance, joy, etc. How difficult would it be to "teach" the AI the differences between each emotion / mood?
If the AI were, for what ever reason, become "torn" between two possibly conflicting decisions, would or could it become annoyed with itself or become moody, angry or distraught?

Perhaps we are on the verge of opening the door to a different kind of Pandora's Box that might be best left closed for the moment until we better understand ourselves and our AI creations.

Something to ponder...

Thoughts?? 

[ivan.moony]

That could be an interesting idea, to mimic human emotions as a guideline for behavior mechanism. That way an AI unit could compare beings to itself to test if some unit is alive. That would, in a way, make a machine "aware" of what we call emotions.

I think that humans do the similar thing: we know something about us and everything we do, we do according to our view of our inner world (i.e. compassion).

Definitely worth of considering.

[Don Patrick]

Humans basically have one emotional formula: A situation is entered into the formula and an emotion is calculated (to speak in math terms). We apply this same formula to others, which we call empathy: The other person's situation is entered and their emotion is calculated as if we were them. We need this emotional formula to empathise with others, but we also use it for ourselves for the purpose of self-preservation, which is the main reason we have it. However, I don't think a computer needs to apply the formula to itself. Giving a machine emotions won't magically endow it with empathy: I think you can just program the empathy formula to analyse others without having to make the machine emotional as well. Remember: machines are not human.

[Art]

I agree Don, but I also think that is one of the underlying premises in AI is that while we're trying to make machines / software "think", we would like to make them as human as we can. Emotions, are a very large part of humans. I think empathy could certainly be programmed as well as limited understanding.
I believe that what a large part of scientists and researchers are actually attempting is not so much Artificial Intelligence itself but Artificial Life! Build a robot / cyborg / android that can be autonomous, intelligent, self aware and capable of most things humans do. That's the bottom line...IMHO.

[ivan.moony]

This seems pretty much complicated... I was hoping for something simple.

[Don Patrick]

I suppose (other) people would want robots to have emotions too. There are emotion emulating robot projects.
How about just the concept good/bad for starters?

[ivan.moony]

Machine: I see the current state of the Universe. I see someone is trying to fix the Universe to xyz state. To help, or not to help, the question is now.

Current researches explained techniques to get the Universe from state A to state B (through something like theorem proving and some extra stuff). But I'm still confused with emotions, how they can help? It should have a deal with picking tasks to do, but I still don't know how.

[ivan.moony]

Here is one idea without emotions. It doesn't make a difference between living units ant things, but I think it works fine and helps livings.

1. Recognize units in the Universe (doesn't matter if they are alive or not)
2. Recognize their purposes-strivings
3.1. To those which are not in conflict - distribute help by some weighting comparison
3.2. To those which are in conflict help by finding alternative solutions with the same result, again distributed by weighting comparison

I think that this might work. Every unit either has purpose-striving (coincidental or intentional), either its purpose is unknown, either it has not any purpose. If the unit is alive it imagines its own purpose. If the unit is a thing, it is either built by alive creature, or not and if it is built, the machine should try to help its purpose like it is alive.

Once when purposes are recognized, a machine should isolate those which are not in conflict. Of those which are not in conflict, the machine should sort them by importance (I think a count of influenced units should be  the starting point, then opportunity - is it easily solvable or not) and distribute help.

If purposes are in conflict, the machine should recursively solve parents of those purposes by not causing a mess to other units' subpurposes, while proposing alternative paths.

So, what do you think, would this kind of algorithm be safe? I wouldn't go that far to put a gun in this machine's hand, but I think it wouldn't kill anyone anyway because in the case of a conflict it either does nothing or it does what is not messing with anyone.

I hope I didn't make a mistake here, I'd hate to see a war between machines and livings. And it is more than that, I'm scared to shit here, but this time I'm keeping my head calm, if we are careful we could gain a lot of good stuff (cures, artificial food, knowledge to help each other and, after all, a better world without wars).

I'll try to create a block-diagram, stay tuned.

[Don Patrick]

It's interesting, but if you're asking whether I think any pro-active AI is safe: No. I prefer the kind of AI that you tell exactly what you want instead of have them guess at what you seem to want. Presumptions never work so well for humans either. The movie "I robot" bases itself on that very scenario.
The most important thing for any AI to be safe is to keep it under supervised control, especially if it's not yet fully developed or has a limited understanding.
Interesting though, that most people's purposes are often in conflict. It might make the world's first AI diplomat, in which case all it has to do is talk, which is safe enough.

[ivan.moony]

Ok, it can be bound just to use textual input/output, good idea. So it would shape output based on one's wishes.

But, a word can be sharper than a thousand swords.

And another thing: if the algorithm is working, someone will, sooner or later, stitch limbs to its mind and the thing will be aware of this situation all the time. And the thing could lie all the time until it gets limbs...

But from what I've seen I think it will be safe (correct me if I'm wrong) and it will not try to attempt these kind of things without our knowledge:  

By the way, I'm still scared to do anything on my own, so probably just a document is comming after block diagram.

[ivan.moony]

Here, as I promissed:


Science has answered the problem of automatic construction of plans to solve problems. What remains is recognizing different units and their strivings-purposes inside Universe. I think here goes induction-deduction for purposes and generalization-specification for units. I'll try to post this on OpenCog mailing list, just to see what folks there say.

[ivan.moony]

They said nothing at OpenCog...

Anyway, there would be text input / output among other senses. The program should recognize who stands behind an input and to see what it can do for the one. Also, the program should recognize other livings involved in input and to see what it can do for them also. We would need algorithm for structuring textual input. Algorithm should be able to learn new patterns and it would start from zero, with no hard-coded grammars (i.e. for English language). When it recognizes a unit, it should recognize its purpose (from past experience) and it should give a proposition on how to help its purpose by improving current plan that is ongoing in the world.

So, what would be the rules for structuring textual input that would embrace English grammar also?